| { |
| "problems": 256, |
| "samples": 512, |
| "acc": 0.2265625, |
| "acc_se": 0.021387020398926954, |
| "cut_share": 0.771484375, |
| "mean_tokens": 26964.421875, |
| "closed_share": 0.228515625, |
| "p_finish": 0.228515625, |
| "p_correct_given_finished": 0.9658119658119658, |
| "p_correct_given_cut": 0.007594936708860759, |
| "no_answer_finished_share": 0.001953125, |
| "loop_end_share_of_cut": 0.9392405063291139, |
| "tail_dup8_cut_mean": 0.8873164556962025, |
| "median_tokens_correct": 6852.0, |
| "by_source": { |
| "gsm8k": { |
| "problems": 128, |
| "samples": 256, |
| "acc": 0.28515625, |
| "acc_se": 0.031217302652994248, |
| "cut_share": 0.7109375, |
| "mean_tokens": 25230.65234375 |
| }, |
| "math": { |
| "problems": 128, |
| "samples": 256, |
| "acc": 0.16796875, |
| "acc_se": 0.028429478161426106, |
| "cut_share": 0.83203125, |
| "mean_tokens": 28698.19140625 |
| } |
| }, |
| "run": { |
| "name": "dev/a1", |
| "model": "qwen35-a1", |
| "data": "dev256", |
| "n": 2, |
| "shards": 1, |
| "est": 13 |
| }, |
| "sampling": { |
| "temperature": 0.6, |
| "top_p": 0.95, |
| "top_k": 20, |
| "presence_penalty": 1.5, |
| "repetition_penalty": 1.0, |
| "max_tokens": 32768, |
| "thinking": true |
| }, |
| "edit_note": { |
| "name": "a1", |
| "spec": { |
| "about": "remove the loop component from the 64 neurons (projection)", |
| "layers": { |
| "15": { |
| "alpha": 1 |
| }, |
| "19": { |
| "alpha": 1 |
| } |
| } |
| }, |
| "built": { |
| "15": { |
| "alpha": 1.0, |
| "n_cols": 64, |
| "dir": "released", |
| "relative_change": 0.01772376149892807 |
| }, |
| "19": { |
| "alpha": 1.0, |
| "n_cols": 64, |
| "dir": "released", |
| "relative_change": 0.013514249585568905 |
| } |
| }, |
| "edited_sha256": { |
| "model.language_model.layers.15.mlp.down_proj.weight": "c6b6b86da9d129eae4ad04d6de3c840686aaf7a9c31292830f866bd408be483e", |
| "model.language_model.layers.19.mlp.down_proj.weight": "a08bd20fe6bcbe3a6574c686df3ceda03b1d8d992feb9b8802bbe7553469de32" |
| } |
| }, |
| "tokens_total": 13805784, |
| "gpu_seconds": 2379, |
| "tok_per_s_per_gpu": 5915 |
| } |
